The Right to be Forgotten in Pruning: Unveil Machine Unlearning on Sparse Models

TL;DR

The paper investigates how data deletion affects pruning topology in sparse neural networks and introduces un-pruning as a practical method to adjust the sparse structure to reflect the data retained after forgetting. It presents a workflow that integrates existing unlearning algorithms into un-pruning, along with a theoretical bound on un-pruning error and new metrics (IoM/IoU/KL) to assess unlearning quality beyond MIA. Empirically, un-pruning closely approximates retraining+repruning across multiple architectures and pruning schemes, with structured pruning generally easier to approximate, while revealing MIA as an unreliable forgetting signal. The work advances trustworthy sparse model deployment by enabling efficient forgetting, and provides code to foster reproducibility and further research.

Abstract

Machine unlearning aims to efficiently eliminate the memory about deleted data from trained models and address the right to be forgotten. Despite the success of existing unlearning algorithms, unlearning in sparse models has not yet been well studied. In this paper, we empirically find that the deleted data has an impact on the pruned topology in a sparse model. Motivated by the observation and the right to be forgotten, we define a new terminology ``un-pruning" to eliminate the impact of deleted data on model pruning. Then we propose an un-pruning algorithm to approximate the pruned topology driven by retained data. We remark that any existing unlearning algorithm can be integrated with the proposed un-pruning workflow and the error of un-pruning is upper-bounded in theory. Also, our un-pruning algorithm can be applied to both structured sparse models and unstructured sparse models. In the experiment, we further find that Membership Inference Attack (MIA) accuracy is unreliable for assessing whether a model has forgotten deleted data, as a small change in the amount of deleted data can produce arbitrary MIA results. Accordingly, we devise new performance metrics for sparse models to evaluate the success of un-pruning. Lastly, we conduct extensive experiments to verify the efficacy of un-pruning with various pruning methods and unlearning algorithms. Our code is released at https://github.com/NKUShaw/SparseModels .

Figures (14)

• Figure 1: (a) Pruning with the original data vs. repruning with the remaining data. The blue blocks indicate pruned parameters. (b) Structured pruning with the original data vs. the remaining data. The blue blocks indicate pruned channels.
• Figure 2: Intersection ratio and unlearning accuracy.
• Figure 3: Structured un-pruning (IoM).
• Figure 4: Unstructured un-pruning (IoM).
• Figure 5: Unstructured un-pruning (IoU).